When every AI agent thinks it’s in charge

One agent writes the code. A second reviews it. A third prepares the release. On a diagram, the arrangement looks reassuringly familiar.

Now imagine the reviewer examining an older version while the developer starts a new approach. The release agent sees passing tests, but they belong to the change that has just been replaced. Each participant has a reasonable account of its own task. Together, they are preparing the wrong release.

This is an illustrative scenario, but the underlying problem is ordinary coordination. Dividing work creates dependencies. Someone—or some shared system—must establish which version is current, who owns the next action, and what has actually been completed.

The handoff is part of the work

A handoff contains more than a summary. It needs a target, an owner, and a condition for proceeding. “Review complete” is ambiguous unless it identifies the version reviewed and the issues still outstanding.

Software teams already have places to record much of this information: issue trackers, code reviews, build results, and release records. An agent system can use those same records. The important question is whether every participant works from a consistent account of the task.

A conversation can help explain a decision. It becomes a fragile substitute for an authoritative work record when several actors depend on it at different times.

The cost of understanding it again

There is another expense that is easy to miss. Each agent may need to read the same requirements and rediscover the same context. A specialist review can be valuable, but the time and computing spent rebuilding that understanding belong in the cost of the task.

Adding agents can also increase waiting. Work that looked parallel in the plan may depend on a decision only one participant can make. Meanwhile, independent edits can create reconciliation work that did not exist before.

A busier factory is not necessarily a more productive one.

The comparison should therefore concern completed work: quality, elapsed time, total cost, and the effort required to resolve disagreements. The number of active agents is an implementation detail.

When a specialist earns a place

A sensible experiment begins with a clear reason for adding another agent. Perhaps the existing process misses a class of security issue, struggles with a particular testing problem, or needs an independent review.

Give the specialist a bounded responsibility and a precise handoff. Compare the results with a simpler arrangement on comparable tasks. Record duplicated work and human intervention along with any improvement.

Sometimes specialisation will help. Sometimes a single agent with better context and stronger checks may be the clearer design. We are interested in the conditions that make each arrangement useful.

A software factory’s organisation should emerge from evidence about the work. A convincing organisation chart is only the beginning of the investigation.